Writing an AI Use Policy for a Small Team: What It Needs to Say

Someone on your team is already using an AI tool for work. They pasted in a customer email to get a reply drafted, or a spreadsheet to get it summarised. Most of the time nothing goes wrong. The risk is that nobody has said what is fine and what is not, so each person decides alone.

A policy fixes that. It does not need a lawyer or a 20-page document for a team of five. It needs to be short enough that people read it and specific enough that they can follow it on a busy Tuesday.

This guide covers the four questions a small-team policy has to answer, and the settings in common tools that decide whether your answers hold up. It is not legal advice. The rules that apply to you depend on your industry and state, so check them with the regulator or an adviser.

Quick answer: a small-team AI policy needs four rules. Say what data may and may not go into an AI tool, naming personal and client information. Say which tools are approved and who approves a new one. Say that a person checks anything that leaves the business, and who. Say when you tell customers they are dealing with AI. Then check the data settings on the tools you already pay for, because the policy is only as strong as those settings.

Start from what the regulator says

The Office of the Australian Information Commissioner (OAIC) publishes guidance on privacy and commercially available AI products. It was published in October 2024 and updated in January 2025. It is worth reading in full, but four points shape a policy.

First, the Privacy Act applies to personal information you type into an AI tool, and to AI output that contains personal information. The OAIC's recommendation is not to enter personal information, and particularly sensitive information, into publicly available generative AI tools.

Second, it tells businesses to check that a product suits the use, and to look at who can access the data you put in, including the developer and third parties.

Third, it asks for transparency: update your privacy policy, and make clear when a public-facing tool such as a chatbot is not a person.

Fourth, it says AI systems are known to produce inaccurate or false results, and asks for human oversight and staff who can assess outputs. It also warns against a "set and forget" approach, so put a review date in your policy.

Those four points map onto the four rules below.

Rule 1: what data goes in

This is the rule that matters most, and the one people break without meaning to. Write it as a short list of what never goes into a tool that has not been approved for it, and what is fine anywhere.

A useful split:

• Fine anywhere: public information, your own marketing copy, generic questions, drafts with names and details removed.
• Approved tools only: internal documents, pricing, supplier terms, anything under a confidentiality clause.
• Never in an AI tool unless the owner signs off: customer or staff personal information, health, financial or identity details, passwords and keys.

Give examples from your own work. "No customer details" is easy to read past. "No pasting a customer's email thread to get a reply drafted, remove the name and address first" is not.

The line between the second and third group depends on which plan and account the tool is on, which is the next section.

Rule 2: which tools are approved

Name the tools people may use for work, on which account type, and who to ask about a new one. A small team usually has one person who decides, often the owner. The rule can be one sentence: "Use the work account for the tools on this list. Ask [name] before trying another."

The point of an approved list is not control for its own sake. It is that the data rule only works if you know where the data goes. Our Concierge can help if you are choosing which tools to approve, and our research tools guide covers the assistants that cite sources.

Review the list on a set date. Plans and data terms change, and the OAIC guidance is clear that a one-off check is not enough.

What the settings say in the tools you already use

Here is what each vendor says about data, from its own pages. Read the current page for your plan before you rely on this, since terms change.

Claude. Anthropic says it does not use inputs or outputs from its commercial products to train models by default, with exceptions such as feedback you send. On Free, Pro and Max accounts, a Model Improvement setting controls whether chats can be used for training, and incognito chats are excluded. The catch in our review is that paid plans cap usage in rolling windows. For a policy, the point is that the account type changes the data answer.

Gemini. Google says that with Keep Activity on, chats are used to provide and improve its services, including training generative AI models, and that a subset is reviewed by humans. Turning Keep Activity off stops future chats being used for training, though Google says chats are still kept briefly. Work and school accounts can be subject to different terms. Our review's catch matches this: business data protections are not what you get on a personal account. A policy line follows: no work data in a personal Gemini account.

Microsoft Copilot. Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train its foundation models, and that Copilot only surfaces content a user already has permission to see. That second point has a consequence for small teams: if a shared folder is open to everyone, Copilot can surface it to everyone. Check your sharing permissions before you switch it on. Microsoft also lets admins decide which agents and third-party models are allowed. Our review's catch is that the version that works across your organisation's data is a paid add-on.

Notion. Notion says customer data is not used to train models by default, that its AI providers are bound by the same commitment, and that Notion AI only uses content a user can already access. It also says that on lower plans providers keep data for a short period before deleting it, while Enterprise has zero data retention options. Our review's catch is that full use of the AI means the Business plan.

ChatGPT. OpenAI's help pages blocked us when we tried to read them, so we are not quoting them. Our review says to skip it if you need business-grade data controls without paying for a team plan. Read OpenAI's current data controls page for your plan before you write ChatGPT into the policy.

The pattern across these is the same. The work or team plan usually carries stronger data commitments than the personal one, and the personal account is where staff most often end up. Your policy should say which account to use.

Rule 3: who checks the output

The OAIC says AI systems can produce false results. Microsoft's own page says responses are not always factual and that users should use their judgment before sending content to others. Treat that as the standard.

A workable rule: anything that goes to a customer, a supplier, a regulator or the public is read by a person before it leaves, and that person is responsible for it. AI drafts. A named human sends.

Be specific about the risky categories: figures, quotes of prices, legal or medical statements, and anything that cites a source. For citations, open the source before you act on the claim. Our guide to research tools has a worked example of how.

Rule 4: when you tell people

There are two audiences.

Customers. If a chatbot or voice agent talks to the public, say it is not a person. The OAIC asks for this, and your privacy policy should mention that you use AI. Our guide to AI voice agents covers what to ask a vendor before you buy.

People in the room. Recording or transcribing calls and meetings has its own consent rules, and they differ by state. Put "tell people before you record" in the policy, and see our note-taker guide for which tools show a visible recorder and which do not.

You might also decide whether staff must say when AI helped with a piece of work. Many small teams choose to, mostly so a reviewer knows to check it harder.

Keep it to one page

A policy that runs past a page will not be read. A structure that fits:

1. Why we have this (two sentences).
2. The data rule, with examples.
3. The approved tools and accounts, and who to ask.
4. Checking output, and who is responsible.
5. When we tell customers and colleagues.
6. Who to tell if something goes wrong, such as the wrong thing pasted into the wrong tool. Make it safe to report. A team that fears blame hides mistakes.
7. Review date.

Walk the team through it once in 15 minutes, and have them tell you what is missing. They know how they use the tools better than you do.

Where to start

• Check the data and training settings on the accounts your team really uses, starting with the tools above. Compare ChatGPT and Claude if you are deciding which to approve.
• Browse the writing tools and productivity tools we have reviewed if staff are asking for something not on your list.
• If you want help choosing tools or setting up a rollout, try the Concierge, or look at AI agencies that do implementation and training.